Jiwei Li applies deep reinforcement learning—a relatively new technique in which neural networks learn by trial and error—to natural-language processing (NLP), the field of computer science in which programs are made to manipulate human languages.
By using deep reinforcement learning to identify syntactic structures within large pieces of text, Li made machines better at extracting semantic information from them. Syntax refers to the grammatical relationship between words, while semantics refers to their meaning.
In written language, words with a close semantic relationship are not always close together on the page. A verb and its object can be separated by a string of adjectives or a subordinate clause, for example. Previous attempts at getting machines to parse natural language often overplayed the importance of proximity, leading to obvious mistakes. Li’s machine-learning algorithms find the grammatical structure of a sentence to get a much more reliable sense of the meaning. They have become a cornerstone of many NLP systems.
Li grew up in China and studied biology at Peking University before moving to the US, where he began a PhD in biophysics at Cornell. But he soon switched fields, turning to NLP first at Carnegie Mellon and then at Stanford, where he became the first student ever to obtain a computer science PhD in less than three years.
Li has also explored other ways to teach artificial intelligence how to spot patterns in linguistic data. In 2014 he and his colleagues correlated Twitter posts with US meteorological data to see how weather affected users’ mood. First he labeled 600 tweets by hand as happy, angry, sad, and so on. He used this labeled data to train a neural network to assess the mood of a tweet and cross-referenced that mood against geolocation data for about 2% of all the tweets published in 2010 and 2011.
His results were not surprising. Moods worsened when it rained; people expressed anger when it was hot. But for Li it was a lesson in how hidden information could be extracted from large amounts of text.
After finishing his studies in 2017, he moved back to Beijing and founded an NLP startup called Shannon.ai, which now has dozens of employees and $20 million in funding from venture capitalists. Li’s company is building on the pattern-matching work demonstrated in the Twitter weather study to develop machine-learning algorithms that extract economic forecasts from texts including business reports and social-media posts.
Li has also applied deep reinforcement learning to the challenge of generating natural language. For him it is the obvious next step. Once you have learned to read, you can learn to write, he says.
Even the best chatbots still make obviously stupid mistakes, spewing out non sequiturs or displaying a lack of basic common knowledge about the world. The longer a conversation, the harder it is for an AI to keep track of what’s been said. Li’s techniques give AI a good grasp of linguistic structure. In a conversation, keeping track of subjects and objects is easier if the syntax of utterances is explicit. For example, given the question “Shall we get started?” a bot might answer “Of course!”—but that response could follow any question. Li’s technique can instead give responses more like “Yes. We’ve got a lot of work to do here,” referencing the content of the original query.